RAG and Graph RAG in marketing: what they are and why they matter

How an AI answers from your documents instead of “from memory”, why classic RAG falls short when the answer lies in the connections, and what a graph adds.

Updated October 6, 2026 · MAPS editorial team

In brief

RAG (Retrieval-Augmented Generation) is the technique that lets an AI model answer from your documents: first it retrieves the passages most relevant to the question, then it generates the answer based on them. Graph RAG adds a graph of entities and relationships — brands, competitors, channels, audiences — and makes it possible to answer even when the answer lies in the connections between pieces of information.

Why it’s needed

On its own, a language model answers with what it learned during training: general knowledge, frozen at a certain date, and none of your research. RAG gives it access to your memory — research, analyses, reports, decisions — and forces it to answer from that. It’s the difference between asking an outside expert and asking someone who has read your entire archive.

How it works

Index. Documents are split into chunks, and each chunk is turned into an embedding: a numerical representation of its meaning.
Retrieve. When you ask a question, the system looks for the chunks closest in meaning.
Generate. The model receives the question and the chunks, and writes the answer, citing where it comes from.

Where classic RAG falls short

RAG finds passages similar to the question. It works well when the answer sits in one place. It works less well when it lies in the connections: “where are we most exposed to Competitor A?” requires combining information spread across different documents — which channels the competitor operates in, which audiences those channels serve, where the brand is weak. No single chunk contains the answer on its own.

What Graph RAG adds

A Graph RAG extracts entities and the relationships between them from documents and keeps them in a graph. When the question names an entity, the system follows its relationships and places them alongside the text: the model can then reason about the connections, not just the passages. A well-built graph can also tell you how many sources support a relationship and since when.

What to ask of a RAG system

Sources. Every answer must say where it comes from.
Dates. Information from a year ago should be presented as such.
The nature of the information. A fact isn’t an analysis, and an analysis isn’t an opinion read on social media.
Gaps acknowledged. If the material doesn’t contain the answer, the system must say so rather than make something up.
Freshness. Whatever you save should enter the memory right away.
Separation. One client’s information must not end up in answers about another.

How MAPS does it

The memory of every MAPS project is an always-up-to-date RAG: everything the project produces or that you upload goes into the memory, and before every question MAPS checks that it is in sync. Each chunk carries the module it comes from, its nature — fact, analysis, opinion — and its date. On top of that there is Graph RAG: a graph built on an ontology designed for marketing and communication, in which every relationship says how many sources support it and from when to when it was reconfirmed. If a piece of information is missing, the answer says so; and with “Compare with the past” the chat tells you what has changed.

An answer from the MAPS chat, with the source and date of every piece of information

How we work →

Frequently asked questions

What’s the difference between RAG and fine-tuning?

Fine-tuning changes the model by training it on new data. RAG leaves the model as it is and passes it the relevant information with every question. For knowledge that changes often, such as market knowledge, RAG is usually the better fit: you update it by saving a document.

Does RAG eliminate hallucinations?

It reduces them a lot, because the model answers from specific sources. It doesn’t eliminate them entirely: you also need rules that force it to cite and to say when information is missing.

What is Graph RAG?

A RAG that, in addition to text chunks, uses a graph of entities and relationships extracted from documents. It makes it possible to answer questions that require connecting scattered information.

Do you need a RAG for each client?

You need each client’s memory to stay separate. When you reason across several clients or countries at once, every piece of information must carry a label showing where it comes from.

Origin and references

Patrick Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (NeurIPS 2020). For Graph RAG: Darren Edge et al. (Microsoft Research), “From Local to Global: A Graph RAG Approach to Query-Focused Summarization” (2024).

Written by the MAPS editorial team with the help of AI, checked and reviewed by the team.

AI & Marketing·8 min di lettura

Vector embeddings: search by meaning, explained simply

What embeddings are, why they let you find information even without the right words, and where they reach their limits.

Read →

AI & Marketing·8 min di lettura

Marketing knowledge base: the memory that doesn’t get lost

Research, analyses, materials and decisions in one place, searchable in plain language: how to build it so it stays useful, and what changes when an AI reads it.

Read →

Methodology·9 min di lettura

Market research with AI: speed without losing depth

What AI does well in market research, where it goes wrong, and the rules for using it without being fooled: sources, dates, the market’s language, estimates flagged as estimates.

Read →